Github user sachouche commented on a diff in the pull request:
https://github.com/apache/drill/pull/1060#discussion_r164532290
--- Diff: exec/memory/base/src/main/java/io/netty/buffer/DrillBuf.java ---
@@ -703,7 +703,18 @@ protected void _setLong(int index, long value) {
@Override
public ByteBuf getBytes(int index, ByteBuf dst, int dstIndex, int
length) {
- udle.getBytes(index + offset, dst, dstIndex, length);
+ final int BULK_COPY_THR = 1024;
--- End diff --
You are right, I'll put more information on this optimization:
- During code profiling, I have noticed that getBytes() doesn't perform
well when called with small length (lower than 1k)
- Its throughput improves as the length increases
Analysis :
- Java exposes two intrinsics for writing to direct memory: putByte and
copyMemory
- The JVM is able to inline memory access (no function call) for putByte
- copyMemory is a bulk API and this internally invokes libc memcpy
(requires function call)
- The rational is that we are willing to incur a function call if the
associated processing is larger than the overhead; this is almost never the
case for small memory accesses.
---
